Super-resolution speckle wavemeter enabled by a tiny convolutional neural network

Artificial intelligence is driving speckle-based wavelength measurement toward higher resolution. However, most reported wavelength resolutions have yet to surpass the minimum tuning interval (MTI) of the reference light sources utilized in experiments. In this study, we develop a compact convolutional neural network, MiniConvNet, for direct wavelength regression, aiming to transcend the hardware's accessible resolution limit. Using only a 10-cm-long multimode fiber, wavelengths separated by 1 pm can be well resolved. When the resolution-to-MTI ratio reduces to 0.5, our method achieves a mean absolute error (MAE) as low as 50 fm and an R-square value of up to 0.9989. At its maximum potential, MiniConvNet can increase the spectral resolution for four times compared to the calibration limit. This breakthrough offers an attractive solution for miniaturized, high-resolution wavemeters that is scalable in spectral regions where high-performance reference light sources are scarce.